Estimating ground-level high-resolution ozone concentration across China using a stacked machine-learning method

Abstract

Accurate estimation of ambient ozone concentrations is key to evaluating its detrimental impacts on human health and the ecological environment. While China has made strides in modeling ambient ozone exposure, nationwide modeling performance and spatial resolution need improvement. To address this, we developed an advanced three-level method, integrating multi-source data as predictors to estimate daily full-coverage ozone concentrations across China, with a 5-km spatial resolution. This modeling method employs subregional division, machine-learning models tailored to each subregion, and a stacking approach for estimate integration. This approach achieved substantial daily scale model performance, yielding R2 [root mean square error] values of 0.94, 0.83, and 0.74 [9.63 μg/m³, 16.59 μg/m³, 20.97 μg/m³] for sample-based, city-based, and temporal 10-fold cross-validation, respectively. Our modeling process identified six subregions for ground-level ozone modeling: Northeast (NE), North China Plain (NCP), Lower and Middle Reaches Plain of the Yangtze River (LRPYR), Southeast (SE), Southwest (SW), and Middle-West (MW). Among multi-source predictors, temperature and humidity are the most important modeling predictors in all subregions. Validation and comparative analyses highlighted the enhancement of overall model performance through subregional division and stacking machine-learning models, particularly for improving predictive accuracy on days without ground-level ozone data. The resultant estimates showed that high ozone exposure in 2020 was concentrated in the spring and summer. Our method and findings hold potential for improving ground-level ozone estimation and advancing ozone-related epidemiological research and environmental management.